This is a program that offloads the computational component of COMCOT to Nvidia GPU. Currently, the speed up achieved by parallelized code on GTX-1060 comparing to serial one on AMD-FX8150 is nearly 200X. The code is still under development to fulfill the full functionality of the original model.
(possible supported version/development environment):
-
gfortran (4.7/4.8)
-
nvcc (/9.2)
-
GNU make (/3.82)
if you wish to use our sample post-processing script. A python intepretor with the following pakages are needed:
-
python(3/3.6)
-
numpy(1.14)
-
matplotlib(2.2)
-
basemap(1.1.0)
- Nvidia-Graphic-Card with Compute Capability (3.5/6.1)
git clone https://github.com/HandsomeAndy/GPU-comcot.git
- Generally, gnu
make
andgfortran
are pre-installed on linux machines. Trymake --version
andgfortran --version
to check if they exist. If not, use package manager to install them.
- Check whether cuda and the card are installed and worked correctly.
nvcc --version
would show the version of nvidia C compiler, andnvidia-smi
list the Nvidia units ready for GPGPU, if the driver and toolkits are all set. here is the full guide to get it work.
- Modify the first few lines in
Makefile
where indicating the location of your cuda library, compute capability, and name of compilers. The exampleMakefile
works on CentOS whose CUDA library is installed byyum
equiped with 6.1 compute capability Nvidia's cards.
make
-
The input file of COMCOT is called
comcot.ctl
. Modify the file to specify a simulation time, time step, fault parameters, and grid settings. Additionally, a corresponding topographic file (available at etopo) is required for grid construction, and the path to the file should be provided incomcot.ctl
. For example, the fresh downloadedcomcot.ctl
takes../etopo_halk2.xyz
as the topographic file and simulate the 2011 Tōhoku Tsunami. -
To run the simulation, simply execute the program
comcot
./comcot
- The program outputs several files at specified steps during runtime. The naming policy of those files follows the original
comcot
, while using binary as files format instead of ASCII for performance reason. The sample scriptplot_dat.py
can be invoked once the simulation begins. It detects outputed files automatically and processes a series of.png
plots with respect to the tsunami waveheight.
python plot_dat.py .
- Here is the 2011 Tōhoku Tsunami propagation animation from those
.png
files (Note: Region in South Pacific ocean was clipped out by the script)
-
There is an closed source OpenMP parallelized version of comcot developed by Tsunami Lab (SCI-link), and the overall speedup versus sigle-core execution time on i5-2500 CPU is shown below.
-
We can find that the maximum speedup on OMP-version is merely at 15x, while GPU distinctively obtain an overall speedup around 138 regardless which CPU model adopted.
-
Taking hardware cost and accessibilities into considerations, a comparison table is listed as the following: (note: gfortran version of comcot: comcot-gfortran)
- Both GTX1060 and Tesla P100 feature similar computing architecture(sm6.1, sm6.0). Unlike GTX1060, howerver, the number of cores and memory bandwidth are respectively 2.8 and 3.8 times higher in Tesla P100. Thus, it is well suited to test our code scalability using these cards. Therefore, the test case were carried out on both models profiled by
nvvp
and results are shown below:
- According to the profiler, the geometric mean regrading speedup over all computation kernels is 2.27. As we may observe,
momts
andmaximum_recorder
kernel obtained a superlinear speedup at 3.21 and 3.24. The reason is that these kernel depend on heavy memory operations and hence performance were limited by the memory bandwidth of GTX1060.
- kernel profile of GPU-comcot running on Tesla P100. Total time consumption: 121 sec
- kernel profile of GPU-comcot running on GTX1060. Total time consumption: 181 sec
GPU version of COMCOT was developed by Tao, Chiu at Tsunami reseach group, IHOS, NCU and the GPU codes are protected under GPL v3.0. The goal of this work is to librate, cooperate ideas with the community and stimulate the development of a tsunami eraly warning system at a relatively low cost. Original COMCOT version can be found at here.